Professional Documents
Culture Documents
2001 - Bathe - The Inf-Sup Condition and Its Evaluation For Mixed Finite Element Methods
2001 - Bathe - The Inf-Sup Condition and Its Evaluation For Mixed Finite Element Methods
www.elsevier.com/locate/compstruc
Abstract
The objective of this paper is to review the general inf±sup condition for mixed ®nite element methods and sum-
marize numerical procedures (inf±sup tests) for the evaluation of the inf±sup expressions speci®c to various problem
areas. The inf±sup testing of a given mixed ®nite element discretization is most important in order to assess its reliability
and solution eectiveness. The problem areas considered are (almost) incompressible analysis of solids and ¯uids,
acoustic ¯uids, the analysis of plates and shells, and the solution of convection-dominated ¯ows. Ó 2000 Elsevier
Science Ltd. All rights reserved.
0045-7949/01/$ - see front matter Ó 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 5 - 7 9 4 9 ( 0 0 ) 0 0 1 2 3 - 1
244 K.J. Bathe / Computers and Structures 79 (2001) 243±252
conditions are ful®lled relatively easily. However, the 2.1. The inf±sup condition using the complete bilinear
applicable inf±sup condition is frequently dicult to form
satisfy. Since an analytical proof to identify whether the
inf±sup condition is satis®ed is frequently out of reach, Consider a general variational problem given in a
numerical inf±sup test procedures have been designed Hilbert space U with a bilinear (symmetric or nonsym-
[1,3±6]. metric) form A /; w de®ned on U U. The ®rst ar-
It is also possible to design a mixed method in which gument in the bilinear form A ; is a solution function
an inf±sup condition is by-passed [7,8]. However, in that and the second argument is a weighting function. We
case, in general, numerical (arti®cial) constants enter the de®ne the following space:
formulation. The choice of these constants is a source of
diculty, even though a mathematical analysis might ou
U uju 2 L2 Vol; 2 L2 Vol; k 1; 2; 3;
give guidelines as to the values to be used. It is more oxk
desirable to not introduce numerical constants in the
u 0 on Su ; 1
formulation and design a method which does satisfy the
applicable inf±sup condition.
Our experience is that if the applicable inf±sup con- where we assume zero Dirichlet boundary conditions on
dition is satis®ed (and of course, the consistency and Su and L2 (Vol) is the space of square integrable func-
ellipticity requirements are ful®lled as well), the ®nite tions in the volume, ``Vol'', of the body considered,
element scheme is reliable and eective and never
``fails''. This is the reason why we endeavor to only use L2 Vol uju is defined in Vol and
®nite element discretization schemes that satisfy the Z
condition. However, the inf±sup condition is a severe u2 dVol kuk2L2 Vol < 1 : 2
requirement and a method not ful®lling the condition Vol
might still be acceptable.
The objective of this paper is to present some latest Given a linear functional (F ; w) from U to R, we have
experiences in the use and numerical evaluation of the for the continuous problem:
inf±sup condition in various analysis areas. In the fol- Find / 2 U such that
lowing sections, we ®rst present the inf±sup condition in A /; w F ; w 8 w 2 U; 3
two general forms, and then show how speci®c expres-
sions are applied to ®nite element formulations for in- where F is the forcing term. For nonzero Dirichlet
compressible analysis, acoustic analysis, plate and shell boundary conditions, the problem statement can of
structures and high Peclet and Reynolds number ¯ows. course also be reduced to Eq. (3) by including the eect
In the presentation, we endeavor to keep the discussion of the boundary conditions in F ; w [1]. We assume that
short and hence do not give details of derivation but there exists a unique solution of Eq. (3).
refer to appropriate references, where these details can The ®nite dimensional subspace of U is de®ned as
be found. We conclude that mixed ®nite element for- follows:
mulations can now be evaluated as shown in this paper
before it is claimed that an eective formulation has ouh
Uh uh juh 2 L2 Vol; 2 L2 Vol; k 1; 2; 3;
been reached. oxk
A brief explanation regarding the notation used
might be helpful. All continuous and ®nite element uh 2 Tn Vol m ; uh 0 on Su ; 4
variables are de®ned in RN ; N 1; 2 or 3 depending on
the problem considered and are always simply presented where Tn Vol m denotes a ®nite element interpolation
as italic symbols. In the de®nition of the functional of order n over the element m of volume Vol m .
spaces used, the given properties apply to each vector The ®nite element solution of Eq. (3) is obtained by
component. However, the ®nite element matrices and solving the following ®nite dimensional problem:
the nodal point solution vectors are given in bold roman Find /h 2 Uh such that
symbols.
A /h ; w F ; w 8 w 2 Uh : 5
This condition ensures the solvability, stability and where h denotes the characteristic element size, k is the
optimality of the ®nite element solution. order of the complete polynomial in the ®nite elements
The inf±sup condition is crucial in establishing the used, and c is a constant, which depends on the problem
quality of the ®nite element approximation. In general, considered (the solution) but is independent of h. Com-
we have the following relation [1,2]: bining the relations (8) and (14), we have the optimal
error bound
kA
k/ ÿ /h kS 6 1 inf k/ ÿ gh kS : 8 kA k
c gh 2Uh k/ ÿ /h kH 1 6 c 1 h c hk ; 15
c
Let us show that the inf±sup condition implies the in-
equality (8). where c is a constant which depends on the problem
From the inequality (7) with gh vh ÿ /h , for any considered and is inversely proportional to c. Since c, kA
vh 2 Uh we have and c are independent of h, so is c .
The formulation that we considered in this section is
A vh ÿ /h ; wh of very general mixed form. We now proceed to consider
ckvh ÿ /h kS 6 sup
wh 2Uh kwh kT a mixed formulation which is more specialized and is
A vh ÿ /; wh A / ÿ /h ; wh encountered widely in applications.
sup
wh 2Uh kwh kT
2.2. The inf±sup condition for the constraint mixed
A vh ÿ /; wh formulation
sup
wh 2Uh kwh kT
kA kvh ÿ /kS kwh kT Consider now the general mixed formulation of a
6 sup
wh 2Uh kwh kT continuous problem in the classical form:
Given two Hilbert spaces U and E, ®nd u 2 U and
kA kvh ÿ /kS ; 9
e 2 E such that
where we used Eq. (6) and that a u; v b e; v f ; v 8 v 2 U;
16
A / ÿ /h ; wh 0 8 wh 2 Uh : 10 b g; u ÿ c e; g 0 8 g 2 E;
Using the triangle inequality, we thus have where a ; , b ; and c ; are bilinear forms, with
a ; and c ; symmetric forms, f ; is a linear form
k/h ÿ /kS 6 k/h ÿ vh kS kvh ÿ /kS and, of course, U and E are the spaces corresponding to
kA the problem considered. The second relation in Eq. (16)
6 kvh ÿ /kS kvh ÿ /kS
c provides of course the constraint in the ``mixed ®elds'' of
the formulation. As in the previous discussion, for
kA
1 kvh ÿ /kS ; 11 simplicity of presentation, we assume that the prescribed
c
boundary conditions on the solution variables are ho-
which proves the inequality (8). mogeneously zero. We also assume that
We may note that we can write Eq. (8) as a v; w 6 ka k v kU k w kU 8 v; w 2 U;
17
k/h ÿ /kS 6 b d /; Uh ; 12 b q; v 6 kb k v kU k q kE 8 v 2 U; q 2 E;
where b is a constant, equal to (1 kA =c), and d /; Uh where we use kkU and kkE to measure the size of the
is the distance between the exact solution / and the ®nite functions in U and E, and ka and kb are continuity
element interpolation space Uh , constants, and that
246 K.J. Bathe / Computers and Structures 79 (2001) 243±252
where k is a vorticity moment and a is a constant of large where c1 are c2 are constants and C corresponds to the
continuous space
value.
This constraint is an additional one to the constraint C v; v 2 L2 Vol; rot v 2 L2 Vol; v s 0 on S
on the pressure included already, see Eq. (25), and the
resulting inf±sup expression is then with s the unit tangent vector on the boundary surface
R R S. Hence, the inf±sup condition is [4]
qh div vh dVol Vol kh rot vh dVol
inf sup Vol ÿ 1h ; Rh rnh ÿ gh
qh 2Qh ; kh 2Kh vh 2Vh kvh kH 1 kqh kL2 kkh kL2 inf sup ch P c > 0 38
1h 2Gh vh 2Vh kvh kV k1h kC0
ch P c > 0;
35 with c independent of h, and
where dhmin stands for the smallest nonzero eigenvalue of Fig. 3. Inf±sup test results for some MITC plate-bending ele-
Dh and rhmax is the largest eigenvalue of Sh . In practice, ments. A clamped square plate using meshes of N by N dis-
torted elements is considered [4].
the semi-norms instead of the full norms are used in Eq.
(41) (based on the Poincare±Friedrichs inequality [1]).
In the second part, we need to show that the C0h - and MITC plate bending elements. (Note that the values of
0
C -norms are equivalent so that in fact the conclusions ch , see Eq. (42), are plotted here).
reached using Eq. (38) with the C0h -norm are also ap-
plicable when the C0 -norm is used. Considering a se- 3.4. General shell conditions
quence of meshes by subdividing the element size h
consecutively, we have An eective shell ®nite element must be applicable to
1h ; vh membrane- and bending-dominated shell problems.
k1h kC0 sup ! k1h kC0 : 43 Ideally, the element shows optimal convergence char-
h=n vh 2Ch=n kvh kCh=n n!1
acteristics in both problem areas. Elements formulated
So that if we consider using pure displacement interpolations are eective in
membrane-dominated situations but ``lock'' when bend-
k1h kC0 ing is encountered. To arrive at a general shell ®nite el-
hn h inf h
; 44
1h 2Ch k1h kC0 ement discretization, a mixed formulation must be used.
h=n
The formulation must preserve consistency for all strain
we can ®nd an estimate for the constant (when n is large) terms, satisfy the ellipticity condition and, considering
the bending-dominated case, satisfy the inf±sup condi-
k1h kC0 tion.
h1 h inf h
: 45
1h 2Ch k1h kC0 For the analysis of general shell problems, the MITC
shell elements are quite eective [1,18]. Let Uh be the
Then if displacement interpolation space (including the dis-
h1 h P h > 0 46 placements and section rotations) and Eh be the strain
interpolation space; then the MITC shell element for-
for some h independent of h, we can conclude that the mulation considering the bending-dominated case can
C0h - and C0 -norms are equivalent. In some cases, it is be written in the classical form of Eq. (16):
possible to prove analytically that h2 h is a lower bound Find Uh 2 Uh and Eh 2 Eh such that
on h1 h [16] but a general numerical procedure is also ÿ AS
given in Refs. [4,16]. A e Uh ; eAS Vh B Eh ; Vh Fb ; Vh 8 Vh 2 Uh ;
An alternative way to proceed is to consider a plate B wh ; Uh ÿ t2 C Eh ; wh 0 8 wh 2 Eh ;
element to be a special case of a shell element and use the 47
inf±sup test considered in the next section.
We should note that the formulation given in Eq. where A ; is the bilinear form containing the bending
(36) is quite general, but the eectiveness of a discreti- strains eAS obtained by tying to the displacements, Eh
zation depends of course on the speci®c interpolation corresponds to the membrane and shear strain compo-
spaces chosen and the reduction operator Rh . The choice nents divided by t2 , t is the shell thickness, B ; is the
of the function spaces for the MITC family of plate bilinear form corresponding to the membrane and shear
bending elements, which satisfy the consistency, ellip- actions, C ; is the bilinear form constraining with
ticity and inf±sup conditions, is discussed in Refs. [1,17] B ; the strain interpolations to the displacements, and
and Fig. 3 shows some inf±sup test results for a few Fb ; is the linear form of the loads divided by t3 . Since
250 K.J. Bathe / Computers and Structures 79 (2001) 243±252
where c is a constant.
The bilinear form B ; evaluates the membrane and
shear strain energies and these can be calculated intro-
ducing
ÿ
e h V h C eAS Wh ; eAS Vh ;
W Th K 51
3.5. High P
eclet number ¯ow kwh k2T wTh T h wh ; 58
Consider a high Peclet number ¯ow problem, in where nh and wh are the nodal values corresponding to nh
which the velocity is prescribed so that the only un- and wh , respectively. Then, as derived in Ref. [6], the inf±
known is the temperature, h. The ®nite element problem sup value ch is obtained by solving the following eigen-
is [1]: value problem for the smallest eigenvalue kmin ,
K.J. Bathe / Computers and Structures 79 (2001) 243±252 251
Acknowledgements
[13] Pantuso D, Bathe KJ. A four-node quadrilateral mixed- [16] Iosilevich A, Bathe KJ, Brezzi F. Numerical inf±sup
interpolated element for solids and ¯uids. Math Models analysis of MITC plate bending elements. Proceedings,
Meth Appl Sci 1995;5:1113±28. 1996 American Mathematical Society Seminar on Plates
[14] Wang X, Bathe KJ. Displacement/pressure based mixed ®- and Shells. Universite Laval, Canada, 1996.
nite element formulations for acoustic ¯uid-structure inter- [17] Bathe KJ, Bucalem ML, Brezzi F. Displacement and stress
action problems. Int J Num Meth Engng 1997;40:2001±17. convergence of the MITC plate bending elements. J Engng
[15] Wang X, Bathe KJ. On mixed elements for acoustic ¯uid- Computat 1990;7:291±302.
structure interactions. Math Models Meth Appl Sci 1997; [18] Bathe KJ, Iosilevich A, Chapelle D. An evaluation of the
7:329±43. MITC shell elements. Comput Struct 2000;75:1±30.